Networking and Data Centers

Cross-Border VPS Speed and Performance Benchmarking Guidelines: Single-Thread, Multi-Thread, iperf3 Stress Tests, and Evening Peak Measurements

VPSMap Editorial · Updated 2026-09-26 · 20 min read · Plain-text version

In international cloud and hosting selection, network performance is often the key bottleneck for availability and response latency. Yet many engineers and administrators who buy instances advertising 1Gbps or faster ports find high speed-test results alongside sluggish production behavior. This guide establishes a standardized, engineering-based, reproducible framework for cross-border VPS network evaluation, clarifying the relationship between port bandwidth, concurrent streams, and route quality. It explains how to avoid misleading multi-thread test results and collect reliable baseline data under demanding evening peak conditions.


One. Separating Key Metrics and Understanding Physical Models in International Network Speed Testing#

When evaluating a cross-border server, download speed alone cannot describe network quality. International public-network transfers typically involve high latency and a large Bandwidth-Delay Product (BDP). Distinguish these core metrics:

Command / configuration
                    ┌── 物理接口带宽 (Interface Capacity) ── 虚拟网卡硬件限速(如 500Mbps、1Gbps、2.5Gbps)
                    ├── 骨干网 QoS 调度与传输优先级 ────── 电信 CN2 GIA/163、联通 9929/4837、移动 CMIN2/CMI
网络质量多维模型 ──┼── 单线程吞吐量 (Single-Thread) ───── 单一 TCP 流承载能力(真实业务、API、Web 加载核心)
                    ├── 多线程并发吞吐 (Multi-Thread) ──── 多连接聚合跑满端口能力(掩盖丢包的常见陷阱)
                    └── 时延抖动与拥塞丢包率 (RTT & Loss) ─ 路由物理距离决定的基础时延与出口拥塞导致的丢包

1. Single-Thread Throughput: The Only Hard Metric for Evaluating Route Quality#

  • Production Workloads Are Fundamentally Based on Single-Stream Interactions: fetching static web resources, remote API calls, database synchronization, SSH terminal responses, and most real-time transfer protocols operate over a single TCP connection or a limited number of connections.
  • TCP Congestion Control and Multiplicative Halving After Packet Loss: traditional congestion-control algorithms such as Cubic reduce the congestion window (CWND) when packet loss is detected. On international links with a high RTT of 150ms~250ms, even sustained packet loss of just 1%~2% prevents a single TCP connection from expanding its window effectively, causing single-thread speeds to fall to only a few megabits.
  • The Dividing Line in Evaluations: standard international routes may saturate the connection with multiple threads off-peak, yet single-thread performance often collapses during evening peaks. In high-quality BandwagonHost data centers, such as DC6/DC9 with CN2 GIA, or on DMIT Premium/Pro high-priority routes, evening-peak single-thread throughput can still remain stable at 100Mbps~300Mbps or higher. Single-thread performance directly reflects the provider's headroom in high-priority bandwidth purchased from Tier-One carriers.

2. The Misleadingly Good Results of Multithreaded Speed Tests#

  • Multithreaded tools, such as Speedtest in its default mode and multithreaded downloaders, split transfers across 8 to 16 concurrent TCP connections.
  • On a degraded route with 5% packet loss, each connection may achieve only 15Mbps, but 16 connections together can still total 240Mbps, creating the false impression that the route is healthy.
  • Benchmark Conclusion: multithreaded tests only verify the physical bandwidth ceiling assigned to the VPS by the data center (Port Burst Cap); single-thread results are what truly represent the end user's experience.

Two. Key Testing Windows and Identifying Network Peak and Off-Peak Periods#

International performance measurements must specify the testing window. Data without a time context has no value for meaningful comparisons.

Testing WindowInternational Submarine Cable and Backbone StatusCarrier QoS PoliciesHow to Interpret Speed-Test Results
Daytime Off-Peak (08:00 - 18:00)Low International Egress Load and Minor Link CongestionLow Packet Loss on Standard Routes Such as 163 and AS4837Use Only as a Reference for Network Adapter and Data Center Physical Limits, Not as a Basis for Plan Selection
Prime Evening Peak (20:00 - 23:30)International Traffic Surges and Consumer Backbone Egress Links Reach CapacityTriggers Multiple Levels of QoS Packet-Drop Throttling, Discarding Large Numbers of Ordinary PacketsPrimary Evaluation Window, providing a realistic test of route priority and high-availability disaster resilience
Overnight Off-Peak Period (01:00 - 06:00)Lowest International Bandwidth UtilizationNo Throttling or Packet Loss, with Stable Peak PerformanceSuitable for Testing the Host's Hardware CPU Performance and Raw NVMe Performance While Minimizing Network Interference

When writing or recording a review report, clearly specify the local start and end timestamps (CST), your local carrier (such as China Telecom, China Unicom, or China Mobile), and your physical connection setup (such as a direct wired gigabit fiber connection).


Three. Pre-Test Environment Preparation and Provider Console Procedures#

Before stress testing, make sure instance networking, firewall rules, and management credentials are ready. Providers differ significantly in instance security policies and console architecture, and changing settings blindly can interrupt connectivity.

1. Login Credentials and Secure Connection Management#

  • DMIT Instance Credential System:
    • Default Security Policy: DMIT cloud instances disable remote root password login by default and strongly require the use of SSH key pairs during initialization.
    • How Key Changes Take Effect: In the DMIT management console, public keys can be managed in the SSH Keys repository and in the instance's Access tab to adjust login credentials. Pay particular attention to the following:After replacing the key in the control panel, you must reboot the instance through the panel (Reboot) before the new key is written to the system and takes effect.
    • Emergency Access Channel: If SSH access is blocked because a local public key was lost or a firewall was misconfigured, open the following in the DMIT console: Console(VNC/Web Console) for out-of-band maintenance and troubleshooting.
  • BandwagonHost Credentials and Control Panel Features:
    • Three Separate Password Systems: BandwagonHost's password system has three independent components: the client-area billing password, the dedicated KiwiVM management console password, and the instance operating system's root password. You must use KiwiVM's own credentials to log into its panel.
    • Out-of-Band Troubleshooting and Network Diagnostics: the KiwiVM panel includes Interactive Console(Interactive Console), allowing you to enter single-user or rescue mode for recovery even if the network driver fails or the system firewall completely blocks SSH.
    • Data Center Migration and IP Status Constraints: BandwagonHost supports self-service data center migration in the client area, but its official knowledge base explicitly states:When the instance's currently assigned IP is on an unavailability blacklist, the console's data center migration (Migrate) feature is restricted. Before large-scale speed testing, first verify IP health through Ping and port scans from your local network.

2. Planning Speed-Test Traffic Within Refund Eligibility Limits#

International iperf3 stress tests and high-speed Speedtest runs can transfer tens of gigabytes (GB) in a very short time. If the current testing is part ofAcceptance Evaluation for a Newly Purchased Plan, pay close attention to the provider's Terms of Service (TOS) and traffic-usage limits for refunds:

  • BandwagonHost Refund Traffic Limits:
    • The official Terms of Service require refund requests for eligible new purchases to be submitted within 30 days.
    • Critical Usage Limits: When requesting a refund, the service's usage during the billing cycleMonthly Traffic Usage Must Remain Strictly Below 10% of the Total Quota. If a plan with a 1000GB allowance uses more than 100GB during initial testing, refund eligibility is lost immediately. Once a refund is processed, the instance and all associated data are permanently destroyed.
  • DMIT Refund Traffic and Time Limits:
    • The official refund policy distinguishes full and partial refunds. For a full refund on an eligible new instance,Must Be Submitted Within 3 Days of Purchase, with Total Data Usage Not Exceeding 30GB.
    • Refund requests after 3 days but within 30 days are eligible only for a partial refund of the remaining value after applicable deductions, with explicit exclusions where no refund is available.
    • Different DMIT series, such as Premium and Eyeball, have different overage policies: some throttle without disconnecting, while others suspend service or apply additional charges. Confirm quota rules on the instance details page in advance.

[!WARNING] A complete 8-thread bidirectional iperf3 stress test, saturating 1Gbps for 60 seconds, consumes approximately 15GB of data. During server acceptance testing, do not leave stress-test scripts running blindly for long periods. Calculate your traffic usage first to avoid exceeding the provider's full-refund usage limit.


Four. Practical Operating Guidelines for Production-Grade Speed-Testing Tools#

For quantitative benchmarking, abandon outdated third-party Python speed-test scripts, whose interpreter performance and single-core overhead can become bottlenecks above 500Mbps, and use natively compiled binary tools instead.

1. Testing Guidelines for the Native Ookla Speedtest CLI Binary#

Install Ookla's official natively compiled Speedtest CLI on the target VPS:

bash
# Ubuntu / Debian 官方软件源部署
curl -s https://packagecloud.io/install/repositories/ookla/speedtest-cli/script.deb.sh | sudo bash
sudo apt-get install -y speedtest

Core Test Parameters and Commands#

  • Single-Thread Testing (Essential Core Metric): Use -u or select single-connection mode to limit Speedtest to one TCP connection:
    bash
    # 强制单线程模式向默认最优节点测速
    speedtest --single
  • Find and Specify Core Backbone Nodes in Mainland China: The default speed-test server is often the geographically closest overseas node, which cannot measure the quality of the route back to China. List and select core hub nodes from China's three major carriers:
    bash
    # 查询当前可用的临近与推荐节点列表
    speedtest -L
    
    # 指定特定测速节点 ID 发起单线程测试(例如指定国内电信或联通骨干节点)
    speedtest -s <Node_ID> --single
    
    # 针对同一节点发起多线程压测(对照分析并发上限)
    speedtest -s <Node_ID>

Criteria for Interpreting the Output:#

  1. Idle Latency and Jitter: Normal trans-Pacific fiber RTT between China and the United States should be between 130ms~170ms. If evening peak jitter consistently exceeds 20ms, the upstream interconnection route is congested.
  2. Low/High Latency (Latency Under Full Load, or Bufferbloat): observe how much ping latency increases under heavy data transfer. High-end optimized routes, such as DMIT Pro or BandwagonHost DC6 China Unicom return routing, typically add less than 30ms under full load.

Five. In-Depth End-to-End Network Stress Testing: iperf3 Operating Guidelines and Tuning#

Speedtest depends on public test servers' performance and outbound allowances, introducing the variable of third-party bandwidth limits. To measure the precise network capacity between your local operations client and the target VPS,iperf3 is an industry standard in network engineering.

Command / configuration
[ 本地测试机 (Client) ] ── (公网端到端长肥管道) ──> [ 境外 VPS (Server) ]
  发起定向单流/多流发包                          监听指定端口 (默认 5201)

1. VPS-Side Service Deployment and Security Configuration#

Deploy the iperf3 server on the target VPS:

bash
# 安装 iperf3
sudo apt-get update && sudo apt-get install -y iperf3

# 以前台监听模式启动,并绑定测试端口(如 5201)
iperf3 -s -p 5201

[!TIP] If a connection times out, check the firewall inside the system (ufw or iptables) allows both TCP and UDP traffic on port 5201. If firewall changes accidentally block your SSH port, use the provider's out-of-band console, such as BandwagonHost KiwiVM Interactive Console or DMIT Console, to log in and reset firewall rules.

2. Standard Benchmark Matrix for the Local Test Client#

After installing the iperf3 client on your local computer (macOS / Linux / Windows WSL), run the following three standardized tests in order:

Method A: Pure Single-Thread TCP Stress Testing During Evening Peak Hours (Baseline)#

bash
# -c: 服务器 IP; -p: 端口; -t: 持续秒数(建议30s); -P 1: 强制单线程; -i 1: 每秒输出一次报告
iperf3 -c <VPS_IP> -p 5201 -t 30 -P 1 -i 1
  • What to Observe: In the output, Retr(retransmitted packets). If retransmissions remain very low (< 50) during evening peak hours and single-thread throughput stays above 80Mbps, the route is not suffering the effects of backbone packet loss.

Method B: 8-Thread Concurrent Throughput Stress Test (Port Capacity Baseline)#

bash
# 评估多连接并发状态下的聚合吞吐能力
iperf3 -c <VPS_IP> -p 5201 -t 30 -P 8 -i 1
  • What to Observe: Observe the combined bandwidth of the 8 streams (SUM) can consistently approach the provider's specified physical speed limit.

Method C: Reverse Transfer Stress Test (Simulating Data Sent from the VPS to Your Local Machine)#

By default, the client sends data to the server, testing local upload / VPS download. To assess website access or file download performance, you must test the VPS's outbound download path to the user (VPS upload to the local machine):

bash
# -R: 反向传输模式 (Reverse mode),由 VPS 向本地客户端发送数据流
iperf3 -c <VPS_IP> -p 5201 -t 30 -P 1 -R -i 1

Option D: Precisely Quantify UDP Jitter and Actual Packet Loss#

TCP can conceal packet-loss details. Sending traffic at a fixed rate in UDP mode directly reveals the link's actual packet loss:

bash
# -u: UDP 模式; -b: 目标设定带宽(如 50M); 观察丢包百分比
iperf3 -c <VPS_IP> -p 5201 -u -b 50M -t 20 -R
  • Evaluation Baseline: premium routes should maintain UDP packet loss below 1% during evening peaks; ordinary backbone routes often rise to 15%~30% packet loss at peak times.

Six. Comprehensive Host Hardware and Base Environment Benchmarking (YABS)#

Network throughput does not operate in isolation. The virtualization platform's CPU clock scheduling, core allocation, and disk I/O speed directly determine network interface softirq and packet-forwarding capacity.

The Industry's Most Authoritative Integrated Benchmarking Tool Is YABS(Yet Another Bench Script), which should be run as soon as the new server is deployed:

bash
# 执行完整 YABS 测试(包含硬件信息、磁盘 FIO、全球多节点网络与 Geekbench 跑分)
curl -sL https://yabs.sh | bash

# 若实例内存低于 1GB,为避免触发 OOM 杀进程,可使用参数跳过高负荷 Geekbench:
curl -sL https://yabs.sh | bash -s -- -g

Engineering Interpretation of Key Output Metrics:#

  1. FIO Disk Performance (4k Random and 1m Sequential Read/Write):
    • 4k Random read/write performance should remain above 20,000 IOPS. During high-frequency cross-border data writes, excessive disk queue buildup can prevent network buffers from flushing to disk promptly, causing packet loss.
    • 1m Sequential read/write: high-quality enterprise NVMe arrays typically deliver read/write speeds between 1,000 MB/s and 2,500 MB/s.
  2. Geekbench CPU Benchmark Score:
    • Single-Core performance is the key metric governing high-throughput TLS/AES encryption and the ceiling for single-thread TCP processing. Modern AMD EPYC processors generally score above 1,300 per core, ensuring that CPU soft interrupts do not saturate at 1Gbps throughput.

Seven. A Framework for Summarizing Test Reports and Aligning with Long-Term Operations#

A rigorous VPS performance and network evaluation should ultimately produce standardized, structured logs to support comparisons and technical decisions.

1. Standard Review Report Recording Template#

markdown
### VPS 实例基准测试记录表
- **服务商与机型**:[填写商户名称及产品线,如 BandwagonHost / DMIT]
- **测试时间窗口**:2026-09-26 21:15 - 22:30 (晚高峰 CST)
- **本地接入网络**:中国联通 1000M 光纤直连 (上海)
- **目标服务器配置**:1 vCPU / 1GB RAM / NVMe
- **系统核心拥塞算法**:BBR (sysctl net.ipv4.tcp_congestion_control)

#### 核心网络实测数据
1. **本地端到端 (iperf3)**:
   - 单线程下载 (-R -P 1): 215 Mbps (重传数: 12)
   - 8 线程下载 (-R -P 8): 910 Mbps (打满端口上限)
   - UDP 丢包率 (50M 压力): 0.3%
2. **Speedtest CLI (国内指定骨干节点)**:
   - 联通核心节点 (单线程): 下行 240 Mbps / 上行 180 Mbps / 抖动 3.2ms
   - 电信核心节点 (单线程): 下行 190 Mbps / 上行 150 Mbps / 抖动 5.1ms
3. **硬件支撑指标 (YABS)**:
   - FIO 4k 混合读写 IOPS: 32,400
   - Geekbench 6 单核评分: 1,420

2. Aligning Test Instance Lifecycles and Billing#

After verifying performance and deciding to retain the instance as a long-term production node, replace the temporary-maintenance mindset of testing with compliant production operations. Pay attention to the providers' differences in billing and data lifecycles:

  • Invoice Generation and Expiration Handling Policies:
    • BandwagonHost: The official knowledge base explicitly states that the systemDoes Not Silently Charge a Linked Credit Card or PayPal Account. If renewal is enabled, an invoice is usually generated 7 days before the service expires. The system automatically applies account credit only if the user has prepaid a sufficient balance. After testing is complete, keep an eye on billing notification emails or add funds to the account to prevent service suspension for nonpayment.
    • DMIT: The Docs billing rules state that monthly renewal invoices are also generated 7 days before expiration. Once nonpayment puts the service into Suspended status,The System Generally Retains Only About a 3-Day Grace Period; if the invoice remains unpaid past the deadline, the instance and its attached disk data are permanently destroyed and cannot be recovered.
  • Network Baseline Retesting Process: Cross-border submarine cables and upstream carrier peering routes are frequently adjusted dynamically. The goal of standardized speed testing is not a one-off screenshot, but to turn the above iperf3 -P 1 and speedtest --single Schedule it as a regular, lightweight monitoring task that avoids high loads, such as collecting a baseline once a week outside business hours with a scheduled script. This provides solid historical data to isolate faults and support service-quality claims if network quality silently deteriorates or upstream routes change.